Deploy and Host Prompt Optimizer on Railway

Prompt Optimizer is an open-source workbench for writing better prompts: paste a prompt, let a model rewrite it with a built-in optimisation template, compare the original and the result side by side, and iterate. This template runs the official web app together with its MCP server, so the same optimiser is available to Claude Desktop, Cursor and any other MCP client.

About Hosting Prompt Optimizer

The template deploys the upstream linshen/prompt-optimizer image, pinned to an exact release rather than latest, as a single service. Inside it nginx serves the web app and a Node process serves the MCP server at /mcp.

It ships with both entrances locked:

Web app behind a password. A random password is generated for the admin user on every deploy. This matters more than it looks: any provider key you set on the service is delivered to the browser inside /config.js, so without a password that key would be readable by anyone who finds the URL. Upstream's own compose file defaults the password to 123456.
MCP endpoint behind a bearer token. Upstream deliberately exempts /mcp from the password (MCP clients cannot do basic auth), and the MCP server has no authentication of its own, so a stock deployment lets anyone call the optimiser on your API key. This template puts an Authorization: Bearer check in front of /mcp using the generated MCP_AUTH_TOKEN. An empty token rejects every request instead of opening the endpoint.

There is no database and no volume. Model settings, history and favourites live in each user's browser (with an optional S3 or WebDAV backup you can configure in the app), and the MCP server is stateless.

Common Use Cases

Prompt engineering for a team: one shared, password-protected URL instead of everyone installing the desktop app or extension
Optimising system prompts: rewrite and test system prompts for agents and assistants before shipping them
MCP tool for coding assistants: call optimize-user-prompt, optimize-system-prompt and iterate-prompt from Claude Desktop or Cursor
Comparing models: run the same prompt against OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter or any OpenAI-compatible endpoint

Dependencies for Prompt Optimizer Hosting

Deployment Dependencies

An API key for at least one model provider, if you want the MCP server or want the web app pre-configured. The web app alone works without one: users can enter their own keys in the browser.

Implementation Details

Signing in to the web app. Open the public URL and sign in as admin with the value of ACCESS_PASSWORD from the service's Variables tab.

Adding a provider key. Set one of these variables on the service and redeploy:

VITE_OPENAI_API_KEY
VITE_ANTHROPIC_API_KEY
VITE_GEMINI_API_KEY
VITE_DEEPSEEK_API_KEY
VITE_OPENROUTER_API_KEY
VITE_CUSTOM_API_KEY together with VITE_CUSTOM_API_BASE_URL and VITE_CUSTOM_API_MODEL for any OpenAI-compatible endpoint (Ollama, vLLM, LiteLLM, Azure and so on)

Keys set here are shared with everyone who knows the web password, because the app runs in the browser and calls providers directly from it. If that is not what you want, leave them empty and let each user enter their own key in the app.

The MCP server needs a key. It picks its default model from the variables above and does not start when none is set. MCP_DEFAULT_MODEL_PROVIDER (for example openai, anthropic or custom) chooses between several.

Connecting an MCP client. Use the Streamable HTTP transport with the URL https://YOUR-DOMAIN/mcp and the header Authorization: Bearer YOUR-MCP_AUTH_TOKEN. For example, in Cursor's mcp.json:

{
  "mcpServers": {
    "prompt-optimizer": {
      "url": "https://YOUR-DOMAIN/mcp",
      "headers": { "Authorization": "Bearer YOUR-MCP_AUTH_TOKEN" }
    }
  }
}

A request without the token, or with a wrong one, gets 401. So does a correct token while the MCP server is not running, which means no provider key is set.

Rotating credentials. Change ACCESS_PASSWORD or MCP_AUTH_TOKEN in the Variables tab and redeploy. Nothing else stores them.

Health. Railway checks /livez, which nginx answers on its own, so a deployment with no provider key is still healthy.

Resources and cost. Measured on Railway: about 70 MB of RAM at idle and around 170 MB while handling MCP calls, with negligible CPU. Expect roughly $1–3 a month on usage-based pricing. Model calls are billed by your provider, not by Railway.

Why Deploy Prompt Optimizer on Railway?

Railway is a singular platform to deploy your infrastructure stack. Railway will host your infrastructure so you don't have to deal with configuration, while allowing you to vertically and horizontally scale it.

By deploying Prompt Optimizer on Railway, you are one step closer to supporting a complete full-stack application with minimal burden. Host your servers, databases, AI agents, and more on Railway.